ZAI’s Atlas 950 cluster is said to use 524,288 Ascend chips
bookwormengr · x · 2026-07-21
The post argues that ZAI has implemented an Atlas 950 SuperCluster at massive scale:
- 524,288 Ascend chips in the full cluster.
- 1 SuperCluster = 64 SuperPods.
- 1 SuperPod = 8,192 NPUs arranged as 8 rows of 16 NPU racks, with 64 NPUs per rack.
The image further breaks down the stack:
- Atlas 950 SuperCluster: 512 EF FP8 / 1,024 EF FP4, 1 EB/s interconnect, 2 EB/s memory bandwidth, 73,728 TB HBM, about 10,240 cabinets and 64,000 m².
- Atlas 950 SuperPod: 8 EF FP8 / 16 EF FP4, 16 PB/s interconnect, 32 PB/s memory bandwidth, 1,152 TB HBM, 160 cabinets.
- Row: 128 TB UB memory domain and 144 TB HBM, with 1,024 × 950DT and 4 PB/s memory bandwidth.
It’s essentially a detailed architecture diagram for a giant clustered AI system, emphasizing how the system scales from racks to pods to the full cluster.
Related event: Z.ai Reportedly Launches 1GW Data Center with Domestic Chips(10 posts)→
More from Infra
- MiniMax H3 runs fully on a Galaxy S26: 5-second video with audio in ~7 minutes — BidFunny4862 · 2026-10-27
- Unconventional AI acquires inference chip firm NeuReality in 1,000x efficiency bid — NaveenGRao · 2026-10-27
- Musk: robot RAM demand will dwarf everything else, and 100% of superintelligence compute goes to deep space — elonmusk · 2026-10-27
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27